Professional Documents
Culture Documents
LECH A. GRZELAKa,b,∗
a Mathematical Institute, Utrecht University, Utrecht, the Netherlands
b Financial Engineering, Rabobank, Utrecht, the Netherlands
Abstract
We focus on extending existing short-rate models, enabling control of the generated implied
volatility while preserving analyticity. We achieve this goal by applying the Randomized Affine
Diffusion (RAnD) method [11] to the class of short-rate processes under the Heath-Jarrow-
arXiv:2211.05014v1 [q-fin.CP] 9 Nov 2022
Morton framework. Under arbitrage-free conditions, the model parameters can be exogenously
stochastic, thus facilitating additional degrees of freedom that enhance the calibration proce-
dure. We show that with the randomized short-rate models, the shapes of implied volatility can
be controlled and significantly improve the quality of the model calibration, even for standard
1D variants. In particular, we illustrate that randomization applied to the Hull-White model
leads to dynamics of the local volatility type, with the prices for standard volatility-sensitive
derivatives explicitly available. The randomized Hull-White (rHW) model offers an almost
perfect calibration fit to the swaption implied volatilities.
Keywords: Randomization, RAnD Method, Randomized Hull-White Model (rHW),
Stochastic Parameters, Stochastic Collocation, Quantization, Short-Rate Models
1. Introduction
Choosing a pricing model, especially in risk management and for Valuation Adjustments
(xVA), is a balancing act between accuracy and pricing speed. Often, the calculation speed
is critical in determining the model of choice, especially when handling large portfolios. The
modern approach for modelling interest rates is based on the Heath, Jarrow, and Morton
(HJM) [13] framework with an arbitrary term-structure of volatility and covariance of forward
rates across maturities. The framework is generic, and essentially any sensible term structure
dynamics can be modelled by the HJM technique. Unfortunately, the number of models giving
rise to analytical formulae for option pricing or even zero-coupon bonds is minimal. Therefore,
models that can generate realistic implied volatilities while allowing for efficient derivative
pricing are in high demand, especially when the portfolios need to be evaluated multiple times
and for different market scenarios due to regulatory pressure.
A family of the HJM models that has gained particular acceptance from practitioners and
academia is the family of affine models [7]. Over many years, various affine models have
been studied, leading to significant contributions regarding efficient pricing, simulation and
calibration (for the overview, see [4]). In particular, the classic affine short-rate models, like
the Hull-White [14] model, are popularised due to the closed formula for zero-coupon bonds
and semi-analytic swaption pricing. On the downside, they lack sufficient flexibility to calibrate
to market-implied volatilities. Additional model flexibility without compromising numerical
efficiency is highly desired. A straightforward approach to incorporate and control the implied
volatility smile and skew into the short-rate models is to define a stochastic or a local volatility
process, for example, in [6, 8, 9]. Although such models give us a closed-form expression for
1
the zero-coupon bond, pricing more advanced derivatives like options often involves advanced
numerical techniques like Fourier inversion or numerical integration, which is undesired for the
pricing of high-volume derivatives. A survey of the most popular stochastic volatility models
in the interest rate world can be found in [1].
This article focuses on extending existing interest rate models with model parameters to
be random via the so-called randomization method (RAnD) [11]. Randomization by making
the model parameters stochastic is a technique for more flexibility in the stochastic model-
thus, improving the calibration quality while preserving the analytic properties of the models
considered. The concept of randomization represents the uncertainty of potentially hidden
states that are not sufficiently captured by deterministic parameters. Conventionally, under
the affine framework, an extension by a stochastic parameter would require the model to meet
linearity conditions - this does not have to be the case under RAnD. The method requires
an affinity to hold, given a specific realization of the stochastic parameter. In other words,
it builds an outer layer over the affine models and allows a stochastic parameter setting in
that layer. The quadrature rule handles the numerical complexity associated with an infinite
number of parameter values and reduces it to only a few critical parameter realizations. The
selection of these points is based on the moments of the stochastic parameter.
The RAnD method is generic and can be applied to any pricing model. For the sake of
simplicity, however, we focus on one of the most popular interest rate models, namely the
Hull-White (HW) model [14]. In particular, we will show that applying the RAnD method to
the HW model results in a local-volatility type of dynamics. The local volatility dynamics may
seem like a “deal-breaker”; typically, the local volatility models, especially in interest rates,
are notoriously tricky to operate, i.e., closed-form option pricing is hardly possible. However,
although the randomized Hull-White (rHW) model is of the local volatility type, the option
pricing is as complex as the standard, unrandomized model. Furthermore, the presented model
offers fast and accurate calibration as the local volatility dynamics provide more flexibility in
generating volatility smiles and skews for various market conditions.
Our method also addresses a challenging problem stated by Brigo and Mercurio in [3], where
the mixture of lognormals was studied. In their paper, it was shown that great flexibility to
standard models like Black-Scholes was introduced by considering a convex combination of
the associated probability density functions: ω1 f1 (x; σ1 ) + · · · + ωN fN (x; σN ) with ω1 + · · · +
ωN = 1 and some σi > 0, i = 1, . . . , N. However, the problem encountered was the large
number of the model parameters that are difficult to interpret and relate to the corresponding
implied volatilities. As stated in their work: “the absence of bounds on the parameter N
implies that a virtually unlimited number of parameters can be introduced in the dynamics
so as to be used for a better calibration to market data. (...), one has to find the correct
tradeoff between model flexibility and number of parameters so as to avoid both poor calibration
and over-parametrization issues”. However, this problem is resolved with the RAnD method.
The method transitions from a continuous randomizer ϑ to a unique sequence of weights ωi
and associated values σi . In this setting, the model calibration is performed by varying the
parameters of the randomizer, ϑ; therefore, the number of free parameters is manageable.
Moreover, as presented in this work, the varying parameters of ϑ enable a transparent control
of the implied volatilities.
This article comprises five sections. First, the details of the HJM framework and pricing
equations using the RAnD method are provided in Section 2, and a discussion on the arbitrage-
free conditions is covered in Section 2.1. In Section 2.2, we consider the rHW model and derive
the dynamics of the randomized stochastic processes. Section 2.3 discusses different variants
of the rHW model parameter choices, namely the randomization using univariate or bivariate
distributions. Section 3 focuses on derivative pricing; in particular, the pricing equations for
options on zero-coupon bonds and swaptions are provided. Numerical results are presented
in Section 4, where in Section 4.1, the swaption implied volatility surfaces for HW and rHW
models are compared. Section 4.2 analyzes the randomizers’ impact on controlling implied
volatilities for swaptions. Next, the calibration results are presented in Section 4.3, and in
Section 4.4, the extension to the bivariate randomization case is established. Finally, the
convergence results are illustrated in Section 4.5 and Section 5 concludes.
2
2. The Heath-Jarrow-Morton models, affinity and randomization
The instantaneous forward rate, f (t, T ), is determined by the volatility driver γ(t, T ), which
can be defined as a deterministic function, stochastic process or by a random variable. The
case where γ(t, T ) is defined as a random variable, thus time-invariant, will be associated with
parameter randomization.
We focus here on the short-rate dynamics under the HJM framework. Suppose that f (0, t),
RT ∂
α(t, T ) and γ(t, T ) are differentiable in their second argument, with 0 ∂t f (0, t) < ∞, then
the short-rate process under the HJM framework is given by,
with Z t Z t
∂ ∂ ∂
ζ(t) = α(t, t) + f (0, t) + α(z, t)dz + γ(z, t)dW (z).
∂t 0 ∂t 0 ∂t
A particular class of popular models are the models belonging to the Affine-Diffusion interest
rate models. It is shown [7] that in this class, for process r(t), the discounted characteristic
function is of the following form:
" !#
Z T
φr (u; t, T ) = Et exp − r(s)ds + iur(T ) = exp (A(u; τ ) + B(u; τ )r(t)) , (2.3)
t
with the expectation under risk-neutral measure Q for τ = T − t, where A(u; τ ) and B(u; τ )
form the fundamental solution to the corresponding pricing PDE and satisfying the complex-
valued ODEs, see the work by Duffie-Pan-Singleton [7]. The affinity conditions are rather strict
as they impose a linear structure on the model, i.e., the model dynamics and its covariance
structure needs to be linear in the state variables. The representation in (2.3) is of particular
importance as it gives an explicit closed-form expression for the integrated short-rate process
and therefore determines the price of a zero-coupon bond (ZCB). Models not leading to an
analytic ZCB price are rarely used for pricing purposes. Unfortunately, because of the affinity
constraints, the model extensions for some model parameters to be stochastic are minimal. In
this article, we focus on relaxing this constraint and letting the model parameters be random
while benefiting from the closed-form solutions for the ZCBs and pricing, as presented in (2.3).
We consider a vector Θ = [ϑ1 , . . . , ϑd ]T , with d ∈ N representing the number of randomized
parameters, where each ϑi is a possibly correlated, time-invariant, random variable. A realiza-
tion of ϑi is indicated by θi , ϑi (ω) = θi , and consequently the realization for Θ is indicated by
θ = [θ1 , . . . , θd ]T . Under the RAnD framework, parameters are not driven by a stochastic pro-
cess but only by a random variable; however, an extension to piece-wise stochastic parameters
is also possible (see [11] for details).
In a nutshell, the RAnD method relies on quadrature integration and conditional expec-
tation. Therefore, the technique can be applied wherever we deal with an expectation. In
principle, randomization can be applied to Equation (2.3), which would provide us with a
randomized ZCB; on the other hand, the randomization can be applied at the valuation level,
where the expectation of the expected payoff is computed. We prefer the latter approach to
construct the randomized variant of any interest rate model by simply evaluating the model
on a particular realization of the model parameters. This will be particularly important when
pricing non-standard derivatives.
In the randomized HJM framework, the volatility in (2.1) may be dependent on random
parameters ϑ1 , . . . , ϑd , i.e., γ(t, T ; {ϑ1 , . . . , ϑd }). Let us now consider a stochastic model un-
der the HJM framework for derivative pricing with the price denoted as V (t, r(t; θ)) with θ
3
indicating a vector-valued realization of the model parameter, Θ. V (t, r(t; θ)) may correspond
to a model of Vašiček, Hull-White, Black-Karasiński or any other model that belongs to the
class of HJM [4]. Then, the pricing under randomized parameter space will be given by the
following relation:
Z
V (t, r(t; Θ)) := Et [V (t, r(t; Θ = θ))] = V (t, r(t; Θ = θ))fΘ (θ)dθ, (2.4)
Rd
where the integration takes place over the parameter space Θ ∈ Rd and fΘ (θ) corresponds
to the associated d−dimensional probability density function. Equation (2.4) shows that a
random model parameter can be considered averaging over possible parameter realizations.
From a numerical perspective, the integration in (2.4) is expensive.
The randomization enables for randomizing d−dimensional space of parameters; however,
for simplicity, we will mainly focus on single or bivariate parameter randomization, ϑ1 (ω) = θ1
and ϑ2 (ω) = θ2 , thus also letting for the correlation between stochastic parameters.
Our aim is twofold: we aim to provide numerically efficient methods for the computation
of the pricing Equation in (2.4) and to determine the stochastic differential equation for the
randomized short-rate processes, r(t; ϑ) where ϑ is the chosen random parameter. The insight
into the dynamics of the randomized stochastic process will help us categorise the randomized
processes and the structure of their drift and volatility coefficients.
Theorem 2.1 (Pricing formula for the randomized model). Consider a random variable ϑ,
defined on some finite domain Dϑ := [a, b], with its PDF, fϑ (x), CDF, Fϑ (x) and a realization
θ, ϑ(ω) = θ such that for some N ∈ N the moments are finite, E[ϑ2N ] < ∞. Let V (t, r(t; ϑ))
be the price of a financial derivative depending on the short-rate r(t; θ) with parameter θ, then
the randomized price, V (t, r(t; ϑ)) is given by:
Z N
X
V (t, r(t; ϑ)) = V (t, r(t; θ))dFϑ (θ) = ωn V (t, r(t; θn )) + N , (2.5)
[a,b] n=1
1 ∂ 2N
N = V (t; r(t, ϑ = ξ)), a < ξ < b. (2.6)
(2N )! ∂ξ 2N
Proof. Define a payoff depending on the short-rate r(t), 0 < t < T , by χ(r, T ), then, for a
stochastic parameter ϑ with the help of conditional expectation, we find:
M (t)
V (t, r(t; ϑ)) = Et Et χ(r, T ) ϑ = θ =: Et [V (t, r(t; θ))] , (2.7)
M (T )
with M (t) being the money-savings account. By application of the quadrature rule, Equa-
tion (2.7) becomes:
Z N
X
V (t, r(t; ϑ)) = V (t, r(t; θ))dFϑ (θ) = ωn V (t, r(t; θn )) + N .
[a,b] n=1
Here, the error is given by N , for a < ξ < b, and the pair {ωn , θn }N
n=1 forms the Gauss-
quadrature weights and the nodes based on the parameter distribution, as proven in [5] (p.180,
Theorem 3.6.24).
Theorem 2.1 gives us an explicit relation between a continuous randomizer ϑ and the
discretization in terms of pairs {wn , θn }N
n=1 , leading to a simplification of the integral in (2.4).
In this paper, these pairs are based on moments of the randomizer, ϑ, and, for completeness,
the detailed computation procedure is given in Appendix A.2.
The associated error N in (2.6) can be interpreted as a cost when moving from the contin-
uous to a discretized random parameter, ϑ → {ωn , θn }N n=1 . To some extent, the pricing with
4
a discrete set of parameter realizations, θn , and the associated probabilities, wn , resembles
the so-called regime-switch method where a finite set of states is defined. However, from a
practical perspective, it is challenging to deal with N pairs of the possible variable states and
the associated probabilities, especially in the calibration procedure.
The procedure described in Appendix A.2, however, simplifies this problem as we can
control the quadrature pairs employing the parameters of the randomizer ϑ. To illustrate this
process, let us consider a stochastic parameter with a randomizer ϑ ∼ N (µ, σ 2 ); thus, the
stochastic parameter, ϑ, is driven by two parameters µ and σ 2 . By application of Appendix
A.2, for any set of parameters µ and σ 2 , we compute the corresponding pairs {ωn , θn }N n=1 ;
therefore, in the calibration procedure we will only vary µ, and σ 2 . This procedure drastically
reduces the number of associated model parameters, thus facilitating the calibration. We also
stress that the set of optimal pairs is solely obtained based on the moments of the randomizing
random variable, ϑ, implying that every random variable with, preferably, closed-form moments
may be used for randomization. In Table A.2, a few selected random variables and their
moments are tabulated. In Section 4.3, we will also classify the randomizers, ϑ, for which the
computations can be greatly reduced or even tabulated.
To emphasize the flexibility of the RAnD method, we focus on the application in the Hull-
White model, where we will apply randomization to each model parameter and consider an
extension to a bivariate case. In the following sections, we consider a finite number of realiza-
tions of ϑ. We will denote them as θ1 , . . . , θN , for some N ∈ N. These “specific” realizations we
will interchangeably call either “collocation” [12] or “quadrature” points. Finally, by ϑ(â, b̂),
we denote that ϑ is a random variable driven by parameters: â, b̂.
with A(τ ; θ) := A(0, τ ; θ) and B(τ ; θ) := A(0, τ ; θ) in (2.3) for τ = T − t, and r(t; θ) indicates
a short-rate model with constant parameter realizations θ. By application of Theorem 2.1, the
randomized ZCB is known explicitly, and it is presented in Corollary 2.1.
Corollary 2.1 (ZCBs under Randomized Affine Jump Diffusion Processes). Let r(t; θ) rep-
resent an affine short-rate process with some constant parameter θ. Assuming that the corre-
sponding ChF, φr(T )|ϑ=θ (·), is well defined and 2N times differentiable w.r.t. θ, the randomized
ZCB is given by,
N
X N
X
P (t, T ; ϑ) = ωn P (t, T ; θn ) + N = ωn eA(τ ;θn )+B(τ ;θn )r(t;θn ) + N
n=1 n=1
=: P (t, T ; {θn }N
n=1 ) + N , (2.9)
with A(τ ; θn ), B(τ ; θn ) being the real-valued functions obtained from Riccati-type of ODEs
available for affine models [16].
Proof. The proof is a consequence of Theorem 2.1.
Representation (2.9) illustrates that the ZCB, P (t, T ; ϑ), randomized with stochastic pa-
rameter ϑ can be expressed as a weighted sum of the ZCBs evaluated at a given realization of
ϑ and where ω1 + · · · + ωN = 1. Note that under the HJM framework, each ZCB, P (t, T ; θn )
in (2.9) is arbitrage-free; therefore, the convex combination is too. This can be shown by
checking whether P (S, T ; {θn }N n=1 )/M (S) is indeed a martingale for some t < S < T :
"P # N
N
P (S, T ; {θn }N
n=1 ) n=1 ω n P (S, T ; θ n ) X P (S, T ; θn )
Et = Et = ωn Et ,
M (S) M (S) n=1
M (S)
5
since every discounted ZCB, P (S, T ; θn )/M (S), is a martingale under the Q measure we have,
N
P (S, T ; {θn }N P (t, T ; {θn }N
X
n=1 ) P (t, T ; θn ) n=1 )
Et = ωn = .
M (S) n=1
M (t) M (t)
This result shows that the application of the RAnD method enables the transition from a
continuous random variable to a discrete one; the cost of this transition is expressed by N .
Moreover, as shown above, the error N in (2.9) does not impact the arbitrage conditions-
the pricing under the RAnD method is simply a weighted average of arbitrage-free prices.
However, from a practical perspective, keeping this error as small as possible is still essential.
This is particularly important when calibrating the randomized model to market data, i.e., the
calibration procedure is associated with varying parameters of the variable ϑ. It is, therefore,
important that the discretized version resembles the continuous version as closely as possible.
In the numerical section, Section 4.5, we will investigate the convergence aspects of different
randomizers and an optimal number of expansion terms, N .
t
η2
Z
µr(t) := r0 e−λt + λ ψ(z)e−λ(t−z) dz, 2
1 − e−2λt .
σr(t) := (2.14)
0 2λ
Moreover, for ψ(t) constant, i.e., ψ(t) ≡ ψ (in this case we deal with the Vašiček model [18]),
we have limt→∞ Et0 [r(t)] = ψ. This means that the first moment of the process converges to
the mean-reverting level ψ, for large values of t.
As a first step, we check the impact of randomization on the probability density function,
PDF, of the interest rate process r(t) in (2.12). Since the HW model belongs to the affine class
6
of processes, we can benefit from the available ChF. Given a stochastic parameter ϑ, the ChF
is given by:
h RT i h h RT ii
φr (u; t, T ) := Et e− t r(s)ds+iur(T ) = Et Et e− t r(s)ds+iur(T ) ϑ = θ .
By application of the quadrature rule in Theorem 2.1 and Appendix A.2, we find:
Z N
X
φr (u; t, T ) = φr|ϑ=θ (u; t, T )dFϑ (θ) = ωn φr(T )|ϑ=θn (u; t, T ) + F
N, (2.15)
[a,b] n=1
and by utilizing the Fourier transform, the PDF of the rHW model reads,
Z N N
1 X X
fr(T ) (x) = e−iux ωn φr|ϑ=θn (u; t, T )du + F
N = ωn fr(T ;θn ) (x) + F
N,
2π R n=1 n=1
where r(t)|ϑ = θn indicates the HW model with a particular realization, θn , of the randomized
variable. The representation above presents the relation between the densities of the rHW
process, r(t), as a convex combination of the HW processes, r(t)|θn with a particular parameter
θn . We also highlight that the error F N in (2.15) may differ from the quadrature error in (2.6).
Because of different randomization choices, we distinguish the following randomization
types and their associated error:
d
XN
fr(T ) (y) + N , for η = ϑ,
d
fr(T ) (y) = ωn fr(T ;θn ) (y) + F
N =: fre(T ) (y) + e
N , for λ = ϑ, (2.16)
d d
n=1
frb(T ) (y) + bN , for η = ϑ1 & λ|η = ϑ2 .
Under the RAnD method, the PDF, fr(T ) (y) in (2.16) is given as a linear combination of normal
densities that depend on different parameter realizations θn , resembling a similar problem as
presented in [3].
In the next section, we associate the PDF given in (2.16) and find the corresponding SDE
for a 1D stochastic representation.
dr(t) = λ(t, r(t))dt + η(t, r(t))dW (t), r(t0 ) = f (0, 0), (2.19)
with
N
X N
X
λ(t, y) = Λn (t, y)λ(ψ n (t) − y), η 2 (t, y) = ηn2 Λn (t, y), (2.20)
n=1 n=1
where:
ωn fr(t;ηn ) (y)
Λn (t, y) = PN ,
n=1 ωn fr(t;ηn ) (y)
has a strong solution whose marginal density is given by the mixture of normal probability
density functions:
N
X
fr(t) (y) = ωn fr(t;ηn ) (y), (2.21)
n=1
PN
where n=1 ωn = 1 for ωn ≥ 0, n = 1, . . . , N with fr(t;ηn ) (x) the PDF of the HW model with
dynamics, given by:
drn (t) = λ(ψ n (t) − rn (t))dt + ηn dW (t), rn (t0 ) = f (0, 0), (2.22)
2
ηn
where rn (t) := rn (t; ηn ) with ψ n (t) = f (0, t) + λ1 f (0, t) + 1 − e−2λt .
2λ2
Proof. The problem we address is the derivation of the drift, λ(t, y), and the volatility function,
η(t, r(t)), in the SDE:
dr(t) = λ(t, r(t))dt + η(t, r(t))dW (t), (2.23)
such that
N
X
fr(t) (y) = ωn fr(t;ηn ) (y), (2.24)
n=1
where fr(t;η) (y) is the PDF of the Hull-White process in (2.17), with parameter η. Under the
HW model dynamics, the so-called linear-growth condition holds, i.e., ηn2 ≤ Cn (1 + y 2 ) for
i = 1, . . . , N. Assuming that the same non-explosion condition holds for (2.23), i.e., η 2 (t, y) ≤
C(1 + y 2 ), uniformly in t, we start by deriving the Fokker-Planck equation for r(t):
∂ 1 ∂2 ∂
η 2 (t, y))fr(t) (y) −
fr(t) (y) = 2
λ(t, y)fr(t) (y), (2.25)
∂t 2 ∂y ∂y
while for each individual interest rate process, rn (t), in (2.22) we have:
∂ 1 ∂2 ∂
ηn2 fr(t;ηn ) (y) −
fr(t;ηn ) (y) = 2
λ(ψ n (t) − y) fr(t;ηn ) (y). (2.26)
∂t 2 ∂y ∂y
8
By substitution of (2.26) and operator linearity,
N N
∂2 X 1 2 X ∂
ωn η fr(t;ηn ) (y) − ωn λ(ψ n (t) − y)fr(t;ηn ) (y) =
∂y 2 n=1 2 n n=1
∂y
" N
# N
1 ∂2 2
X ∂ X
η (t, y) ω f
n r(t;ηn ) (y) − λ(t, y) ωn fr(t;ηn ) (y).
2 ∂y 2 n=1
∂y n=1
for some time-dependent functions C1 (t), C2 (t) and C3 (t). Since the LHS and RHS of the
equations above need to satisfy uniform convergence requirements implying that for y → ∞,
they converge to 0, the following needs to hold: C1 (t) = C2 (t) = C3 (t) = 0, ∀t. Therefore, the
expressions for λ(t, y) and η(t, y) read:
PN 2
PN
n=1 ωn ηn fr(t;ηn ) (y) ωn λ(ψ n (t) − y)fr(t;ηn ) (y)
2
η (t, y) = PN , λ(t, y) = n=1 PN . (2.27)
n=1 ωn fr(t;ηn ) (y) n=1 ωn fr(t;ηn ) (y)
By setting
ωn fr(t;ηn ) (y)
Λn (t, y) = PN for n = 1, . . . , N, (2.28)
n=1 ωn fr(t;ηn ) (y)
we can write
N
X N
X
λ(t, y) = Λn (t, y)λ(ψ n (t) − y), and η 2 (t, y) = ηn2 Λn (t, y). (2.29)
n=1 n=1
PN
Finally, by taking η∗ := maxi=1,...,N ηn and since n=1 Λn (t, y) = 1, ∀y, we have:
N
X N
X
η 2 (t, y) = ηn2 Λn (t, y) ≤ η∗2 Λn (t, y) = η∗2 = C.
n=1 n=1
Since the volatility parameter η 2 (t, y) is bounded by a constant, the uniform convergence
criterion is satisfied. The uniqueness of the strong solution follows from Theorem 12.1 in [17]
while in [2] (Theorem 2.1), a proof for a generic case for a normal mixture is provided.
Proposition 2.1 illustrates that under the randomized volatility parameter for the HW
model, the normal mixture dynamics resemble a one-dimensional local-volatility-type diffusion
process. The local volatility function is expressed as a weighted volatility squared, ηn2 , of the
constituent processes, rn (t), and where the weights are functions of the quadrature coefficients,
ωn , and the corresponding PDFs, fr(t;ηn ) (y).
Following the same strategy, we derive the dynamics of a process of the rHW model with
the randomized mean-reversion parameter, λ. Proposition 2.2 provides the details.
Proposition 2.2 (Dynamics of the HW model with randomized mean-reversion parameter,
λ). Let us assume a sequence of positive constants λm , m = 1, . . . , M , then the SDE
de
r(t) = λ(t,
e re(t))dt + ηe(t, re(t))dW (t), re(t0 ) = f (0, 0), (2.30)
9
has a strong solution whose marginal density is given by the mixture of normal probability
density functions:
M
X
fre(t) (x) = $m fre(t;λm ) (x), (2.32)
m=1
PM
where m=1 $m = 1 for $m ≥ 0, m = 1, . . . , M , with fre(t;λm ) (x) is the PDF the HW model
whose dynamics is given by:
rm (t) = λm (ψm (t) − rem (t))dt + ηdW (t), rem (t0 ) = f (0, 0),
de
η2
1
1 − e−2λm t .
where ren (t) := ren (t; λn ) with ψm (t) = f (0, t) + λm f (0, t) + 2λ2m
2.3.1. Dynamics of the rHW model with bivariate distribution for λ and η.
An extension of the RAnD method is to consider both HW model parameters random
and follow a bivariate distribution. Such an extension may benefit from the interconnection
between model parameters and possibly the correlation coefficient; however, it would require
that the conditional moments are known explicitly. This is troublesome because only for a few
random distributions the moment functions are known in the closed form. However, if we stay,
for example, within the Gaussian world, such an extension to the 2D case is possible.
Corollary 2.2 (Random parameters with bivariate distribution). Under a bivariate distribu-
tion Θ = [ϑ1 , ϑ2 ] with ζ(ϑ1 ) = {ω1,n , θ1,n }N M
n=1 and conditioned on ζ(ϑ2 |ϑ1 ) = {ω2,m , θ2,m }m=1 ,
the randomized ChF is given by:
N
X M
X
φX (u; t, T ) = ωn ωm φX|ϑ1 =θn ,ϑ2 =θm (u; t, T ) + bN,M , (2.33)
n=1 m=1
where N and M indicate the number of expansion terms for ϑ1 and ϑ2 |ϑ1 respectively, ϑ2 |ϑ1
indicates a conditional random variable, bN,M is the corresponding aggregated error, and the
remaining specification follows Theorem 2.1.
The PDF computation for the bivariate case requires the sequential computation of the
associated weights and the corresponding points, which can be established utilizing Proposi-
tions 2.1 and 2.2. In the first iteration step, we compute the grid points associated with the
volatility parameter, η, and for each realization ηn , we establish the corresponding conditional
PDF. By summing over all possible pairs, the unconditional PDF for the rHW model can be
computed,
M
X N X
X M
fr(T ;ηn ) (x) = $m,n fr(T ;ηn ,λm ) (x), fr(T ) (x) = ωn $m,n fr(T ;ηn ,λm ) (x), (2.34)
m=1 n=1 m=1
where $m,n is defined in (2.32), ωn is given in (2.21) and fr(T ;ηn ,λm ) (x) indicates the PDF of
the HW model with parameters ηn and λm , and is defined in (2.14). Proposition 2.3 provides
the dynamics of the associated rHW model.
10
Proposition 2.3 (Local volatility process for the HW model with randomized parameters).
Let us assume a sequence of positive constants ηn , n = 1, . . . , N , and λm,n , m = 1, . . . , M ,
then the SDE:
db
r(t) = λ(t,
b rb(t))dt + ηb(t, rb(t))dW (t), rb(t0 ) = f (0, 0), (2.35)
with
N X
X M N X
X M
ηb2 (t, rb(t)) = b n,m (t, y)ηn2 , λ(t,
Λ b rb(t)) = b n,m (t, y)λm,n (ψm,n (t) − y),
Λ
n=1 m=1 n=1 m=1
where
ωn $m,n fr(t;ηn ,λm ) (y)
Λ
b n,m (t, y) = P
N PM ,
n=1 m=1 ωn $m,n fr(t;ηn ,λm,n ) (y)
has a strong solution whose marginal density is given by the following mixture of normal prob-
ability density functions:
N X
X M
frb(t) (y) = ωn $m,n fr(t;ηn ,λm,n ) (y),
n=1 m=1
PN PM
where n=1 ωn = 1, m=1 $m,n = 1 for ωn , $m,n ≥ 0, n = 1, . . . , N , m = 1, . . . , M with
fr(t;ηn ,λm,n ) (y) being the PDF of the HW model whose dynamics are given by:
drm,n (t) = λm,n (ψm,n (t) − rm,n (t))dt + ηn dW (t), rm,n (t0 ) = f (0, 0),
2
ηn
1
1 − e−2λm,n t .
and rn,m (t) := r(t; ηn , λm ) with ψm,n (t) = f (0, t) + λm,n f (0, t) + 2λ2m,n
This section focuses on the pricing under the rHW model. The presented pricing equations
utilize Theorem 2.1. Throughout the section, we denote by VHW (t, r(t; θ)) with θ ∈ {η, λ}, the
Hull-White price of a derivative depending on the short-rate r(t; θ) with parameters λ and η,
by VrHW (t, r(t; ϑ)) we denote the pricing under the rHW model.
11
where,
M (t)
VχZ (t, T, S, K; θn ) = EQ
t max(χ(P (T, S; θn ) − K), 0)
M (T )
= χP (t, S)FN (0,1) (χdn ) − χKP (t, T )FN (0,1) (χ(dn − σ̄n )), (3.2)
with P (t, S) and P (t, T ) the ZCB computed from the associated yield curve; P (T, S; θn ) is the
ZCB obtained from the model with the randomized parameter, ϑ; χ = 1 and χ = −1 corresponds
to call and put options, respectively. Depending on the randomized parameter, we have:
ηn2
for η : σ̄n2 = 1 − e−2λ(T −t) B 2 (T, S; λ), (3.3)
2λ
η2
for λ : σ̄n2 = 1 − e−2λn (T −t) B 2 (T, S; λn ), (3.4)
2λn
with
1 P (t, S) σ̄n 1
dn = log + , B(T, S; λ) = 1 − e−λ(S−T ) ,
σ̄n P (t, T )K 2 λ
where the HW model is defined in (2.12) and FN (0,1) (·) corresponds the standard normal CDF.
Proof. The proof is a direct consequence of combining Theorem 2.1 with the pricing of options
on a ZCB under the HW model (as given in [4, 16]).
Now, we consider an interest rate swap with a fixed rate, K, and payment times T =
{Ti , Ti+1 , . . . , Tm } and the corresponding reset rates {Ti−1 , Ti+1 , . . . , Tm−1 } with the payoff
given by:
m
X
Swap
HP/R (T , K) = ᾱ τk (`(Tk−1 ; Tk−1 , Tk ) − K) ,
k=i
for τk = Tk − Tk−1 , with P indicating a swap payer for ᾱ = 1 and a swap receiver for ᾱ = −1
and where `(t; Tk−1 , Tk ) stands for the libor rate over the period [Tk−1 , Tk ] observed at time t.
To determine today’s value of the swap, we evaluate the corresponding expectation of the
discounted future cash flows, i.e., each payment which takes place at the time points, Ti , . . . , Tm ,
needs to be discounted to today,
m m
Swap
X M (t) X
VP/R (t, K, T ) = ᾱ τk EQ
t `k (Tk−1 ) − K = ᾱ τk P (t, Tk ) `k (t) − K , (3.5)
M (Tk )
k=i k=i
where the expectation, ETt [·], is taken under the T −forward measure and where the price of a
swap price at T = Ti−1 is given by:
" m
# m
Swap
X X
VP/R (T, K, T ) = ᾱ 1 − P (T, Tm ) − K τk P (Ti , Tk ) = ᾱ − ᾱ ck P (Ti , Tk ),
k=i k=i
12
Up to this point, the pricing equations do not depend on specific model choices. Now,
however, we consider pricing under the rHW model, which, via the conditional expectation
approach, as presented in Lemma 2.1, yields:
" " m
##
X
Swpt T T
VP/R (t, T, K, T ; ϑ) = ᾱP (t, T )Et Et max 1 − ck P (Ti , Tk ), 0 ϑ = θ (3.6)
k=i
N
" m
#
X X
= ᾱP (t, T ) ωn ETt max 1 − ck P (Ti , Tk ; θn ), 0 + N .
n=1 k=i
Here, N is the associated quadrature error defined in (2.6). Since the inner expectation
in (3.6) resembles the expression for a swaption under the classical HW, we follow the standard
procedure and apply the so-called Jamshidian trick [15] to exchange the maximum operator
and the expectation. The resulting pricing equations for swaptions under the rHW model are
given in Lemma 3.2.
Lemma 3.2 (Pricing of Swaptions under randomized Hull-White model). Consider the rHW
model, with parameters {λ, η} and the randomizing random variable ϑ, which randomizes either
of the model parameters. For a unit notional, a constant strike, K, option expiry T = Ti−1
and a strip of swap payments T = {Ti , . . . , Tm }, with Ti > Ti−1 and accruals τi = Ti − Ti−1 ,
the prices of swaption payer and receiver, P/R := Payer/Receiver, are given by:
N
X m
X
Swpt
VP/R (t0 , T, T , K; ϑ) = ωn ck VχZ (t0 , T, Tk , K̂k (θn ); θn ), (3.7)
n=1 k=i
with a swaption payer, P, for χ = −1, swaption receiver, R, with χ = 1, where VχZ (·) is defined
in (3.2) and where the strike price K̂k (θn ) = exp (A(T, Tk ; θn ) + B(T, Tk ; θn )rn∗ ). Here, rn∗ is
determined by solving, for each parameter realization θn , the following equation:
m
X
1− ck exp A(T, Tk ; θn ) − B(T, Tk ; θn )rn∗ = 0, n = 1, . . . , N, (3.8)
k=i
where
P (0, Tk ) η2
1 − e−2λT B 2 (T, Tk ; {λ, η}) ,
A(T, Tk ; {λ, η}) = log + B(T, Tk ; {λ, η})f (0, T ) −
P (0, T ) 4λ
1
B(T, Tk ; {λ, η}) = 1 − e−λ(Tk −T ) ,
λ
with ck = Kτk for k = i, . . . , m − 1, cm = 1 + Kτm and where the pairs {ωn , θn }, n = 1, . . . , N ,
are based on the randomizer ϑ and computed using Appendix A.2.
Under the rHW model, the pricing Equation (3.7) for swaptions is a direct application
of Theorem 2.1. The pricing under randomized parameters is simply an average of non-
randomized prices, with varying model parameters, accompanied by weights based on the
randomizer.
From the computational perspective, the swaption pricing in (3.7) requires N swaption
prices for different realizations θn , n = 1, . . . , N ; therefore, N optimization problems in (3.8)
need to be solved. This computational complexity can be greatly reduced by employing the
multi-d Newton-Raphson algorithm to determine optimal rn∗ for n = 1, . . . , N.
Remark 3.3 (Computation of sensitivities). The computation of sensitivities under the RAnD
method is straightforward, i.e., because the pricing is expressed as a convex combination, the
sensitivities are expressed as a weighted sum of individual derivatives. For example, in the case
of the swaption pricing, the sensitivity to a particular market quote q is expressed as a sum of
sensitivities of individual options on ZCBs:
N M
∂ Swpt X X ∂
VP/R (t0 , T, T , K; ϑ) = ωn ck VχZ (t0 , T, Tk , K̂k (θn ); θn ), (3.9)
∂q n=1
∂q
k=i
13
with the specification as given in Lemma 3.2.
On the other hand, the sensitivity to the quadrature pairs, {ωn , θn }, ∂θn /∂â and ∂ωn /∂â
may be, for some specific cases, computed analytically (see Section 4.3). Still, in a generic
setting, it is recommended to compute these derivatives numerically, with, for example, finite
differences:
ζ(ϑ(â + δâ )) − ζ(ϑ(â − δâ )) ∂ωn ∂θn
≈ , ,
2δâ ∂â ∂â
where ϑ(â) indicates the dependence of the random variable and parameter â, δâ is the shock
size and ζ(ϑ) : R → {ωn , θn }N n=1 is defined in Appendix A.2. Due to the applied finite
difference shocks to â, an additional bias will be introduced. We expect, however, this error to
be of acceptable magnitude, as is commonly observed in current financial practice.
4. Numerical experiments
In this section, several pricing experiments will be performed. First, we present a detailed
analysis of the rHW model in realistic pricing scenarios. In the first experiment, we explore
the implied volatility smile evolution in time, comparing the implied volatility surface from the
HW and rHW models. As the next step, the study of the parameter randomization on shapes
of implied volatilities will be illustrated. Finally, the calibration results with market data will
be presented in the conclusive experiment. This section will end with numerical experiments
involving bivariate distribution for the model parameters.
VB (T, K, F0 , σs , ᾱ, s) = ᾱ · (F0 + s) · FN (0,1) (ᾱd1 ) − ᾱ(K + s)FN (0,1) (ᾱd2 ), (4.1)
1
√ log(F0 + s)/(K + s) + 1/2σs2 T ,
d1 =
σ T
√
d2 = d1 − σs T ,
with shift parameter s, F0 being the forward rate, σs is the corresponding volatility coefficient,
K being the strike, and T corresponds to the time to option expiry.
Then, the shifted Black’s formula for swaptions is given by:
m
B,Swpt
X P (t0 , Ti−1 ) − P (t0 , Tm )
VP/R = VB (T, K, S(t0 ), σs , ᾱ, s) τi P (t0 , Ti ), S(t0 ) = Pm , (4.2)
i=1 i=1 τi P (t0 , Ti )
with swaption payer, P , for ᾱ = 1 and swaption receiver, R, for ᾱ = −1 and where S(t0 ) is
the corresponding swap rate. The shift parameter, s, is typically perceived as the lower bound
for the interest rates by the market participants and varies depending on the currency. It is
important to note that when inverting the Black’s formula in (4.2), the corresponding implied
volatility σs is a function of the shift parameter, i.e., for different choices of s, the implied
volatilities are different. In principle, this is not a problem as long as the implied volatilities
from the market and model volatilities are computed with the same shift coefficient.
In the first experiment, we compare the implied volatility surfaces of the HW model and
the rHW model. In the rHW model, the speed of mean-reversion is randomized by a uniform
distribution, λ ∼ U([â, b̂]), on an interval [â, b̂]. The numerical results are illustrated in Figure 1.
The results demonstrate that the HW model can only generate implied volatility skew. At the
same time, the randomization of the mean-reversion parameter, λ, shows implied volatility
skew and smile. Although the randomization is not time-dependent, i.e., the parameters are
stochastic but stationary, we observe a time evolution of the implied volatilities. The same
phenomenon has been observed in [3, 11], where the randomized Black-Scholes model was
14
Figure 1: Swaption volatility evolution for the HW and rHW models implied by the
shifted Black’s model. The simulation was performed for varying swaption option expiry,
T , and a fixed tenor of 1y. The parameters specified in the experiment are: for the HW
model: η = 0.005, λ = 0.001 and for the rHW model: η = 0.005 and λ ∼ U ([−0.15, 0.6]).
In the experiment, the implied volatilities are computed with zero shift parameter, s = 0.
Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
70 70
60 60
50 50
Implied Volatility, [%]
40 40
30 30
20 20
10 10
0 0
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)
of the volatility parameter η has a pronounced effect on the level of implied volatility with
minimal effect on the smile. The curvature is visible for uniform randomization. It is essential
15
to note that a higher variance for the normal distribution may give rise to a higher curvature
but may also cause issues related to negative volatilities. Our experiments have shown that
even for distributions defined in the positive domain, the impact on the smile is limited, even
for fat-tailed random variables.
A much richer spectrum of implied volatility shapes is obtained when the randomization
technique is applied to the mean-reversion parameter λ. Figure 3 presents the randomization
with either uniform or normal random variables. A substantial amount of curvature can be
generated by taking the mean-reversion random. We also report that the curvature change
affects the overall volatility level, i.e., it is impossible to keep the level fixed and only adjust the
smile. However, the implied volatility level can be fixed by adjusting the volatility parameter
η. This strategy will be discussed further in the context of model calibration.
Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
20 18
18 16
16 14
Implied Volatility, [%]
12 10
10 8
8 6
6 4
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)
In the final experiment of this section, we consider the randomization of λ ∼ U([â, b̂]) using
uniform distribution and check how the parameters â and b̂ affect the implied volatilities. In
Figure 4, the results show an interesting pattern: the curvature level is mainly driven by the
distance |b̂ − â|, i.e., the larger the distance, the more implied volatility smile is generated.
Changes of either of the parameters affects the volatility level; therefore, some of the volatility
effect, η, can be offset by the interval [â, b̂].
Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
18 18
16
16
14
14
Implied Volatility, [%]
12
12
10
10
8
8
6
6 4
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)
Given the numerical results presented above, we conclude that the most considerable impact
on implied volatilities comes from the mean-reversion, λ, and not from the volatility parameter,
η. This is particularly interesting when we confront these results with Proposition 2.1 and
Proposition 2.2, where the dynamics of the corresponding short-rate processes were derived
16
and where it was shown that randomization of the mean-reversion did not lead to the local
volatility type of dynamics. In contrast, the randomization of η does, which can be explained
by the nature of interest rate derivatives, where derivative prices are driven by the dynamics
of the ZCBs, but not directly by the short-rate process. However, since under the HW model,
the volatility of the ZCBs is given by both parameters (see [16]), the randomization of either
of them will imply a local volatility type of dynamics for the ZCBs.
Table 1: Calibration of the HW and rHW model: parameters determined in swaption calibra-
tion.
The calibration fit is presented in Figures 5, 6 and 7. We report an excellent calibration fit
for all considered option expiries, varying from 1y to 20y. We have used two parameters in all
the calibration cases, just as in the HW model. The results confirm that the RAnD method has
great potential for improving existing pricing methods, even with the same number of degrees
of freedom.
17
Calibration results for: 1Y-1Y Calibration results for: 2Y-1Y
28 30
Market Market
Hull-White Hull-White
randomized Hull-White 28 randomized Hull-White
26
ATM ATM
26
Implied Volatility [%] 24
22
22
20
20
18
18
16 16
0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05 0.055
strike, K strike, K
Figure 5: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 1y and T = 2y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.
20
Implied Volatility [%]
22
18
20
16
18
14
16
12
14 10
12 8
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0 0.02 0.04 0.06 0.08 0.1 0.12
strike, K strike, K
Figure 6: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 5y and T = 10y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.
20 20
15 15
10 10
5 5
0 0.05 0.1 0.15 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16
strike, K strike, K
Figure 7: Calibration results of the HW and the rHW models. The market implied
volatilities for swaptions were obtained on 18/08/2022 for the USD market. Option
expiry: T = 15y and T = 20y and the implied volatility shift: s = 1%. Calibrated
parameters are presented in Table 1.
18
4.4. Pricing under bivariate distributions for the model parameters
This section extends the rHW model and considers a bivariate distribution for both model
parameters. Under a bivariate distribution Θ = [ϑ1 , ϑ2 ] with ζ(ϑ1 ) = {ω1,n , θ1,n }N
n=1 and
conditioned on ζ(ϑ2 |ϑ1 ) = {ω2,m , θ2,m }M
m=1 , the randomized prices are:
N
X M
X
VrHW (t, r(t); ϑ1 , ϑ2 ) = ωn ωm VHW (t, r(t); ηn , λm ) + bN,M , (4.3)
n=1 m=1
where N and M indicate the number of expansion terms for ϑ1 and ϑ2 |ϑ1 , respectively, ϑ2 |ϑ1
indicates a conditional random variable, bN,M is the corresponding aggregated error. The
remaining specification follows Theorem 2.1.
As an example, let us consider a bivariate normal distribution for the pair (ϑ1 , ϑ2 ) with the
corresponding realizations (η, λ) for which we have:
σλ 2 2
ϑ2 |ϑ1 = ηn ∼ N µλ + ρ(ηn − µη ), (1 − ρ )σλ , (4.4)
ση
where ρ is the correlation coefficient between ϑ1 and ϑ2 . In Figure 8, we illustrate the impact
of the correlation coefficient ρ on the swaption implied volatilities. We report that similar to,
e.g., the Heston model, the correlation controls the implied volatility skew. Furthermore, we
observe that a higher positive correlation generates more skew, while a negative correlation
generates more volatility curvature (smile).
Implied Volatilities for RAnD Hull-White Model,T=5 Implied Volatilities for RAnD Hull-White Model,T=5
12 18
17
11
16
15
10
Implied Volatility, [%]
14
9 13
12
8
11
10
7
9
6 8
-1 -0.5 0 0.5 1 -1 -0.5 0 0.5 1
Strike, log(K/F) (log-moneyness) Strike, log(K/F) (log-moneyness)
Remark 4.1 (Volatility term structure and feasible strategy for model calibration). Com-
monly, the volatility-parameter η is piece-wise constant, so the ATM volatilities are well cal-
ibrated. This strategy will also work with the randomized mean-reversion parameter, i.e., the
mean-reversion parameter can be used for smile/skew calibration, while the volatility parameter,
η, will ensure a proper fit of the ATM level.
-10
-10
-15
-15
-20
-20
-25
-25
-30
-30
-35
-35 -40
2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 11
N N
Figure 9: Convergence results for randomization of volatility parameter, and mean re-
version parameter η and λ under the HW model respectively. The base parameters in
the experiment were η = 0.00625 and λ = 0.002.
5. Conclusion
In this paper, we have applied the randomization technique to enhance the flexibility of
short-rate models for interest rates. We have shown that the model parameters driven by a
random variable, instead of being deterministic, facilitate a practical extension of standard,
well-popularized models. In addition, this article points out how the normal mixture (a sum
of the Hull-White PDFs) can be expressed as a one-dimensional diffusion process with a local
volatility function.
We have illustrated that, for the randomized Hull-White model (rHW), one can utilize
the available closed-form pricing equations and benefit from flexibility in controlling implied
volatilities. In particular, we have shown that the rHW model results in almost perfect swaption
calibration.
Finally, the RAnD method is generic and is not limited to any particular modelling choice;
therefore, it opens many possibilities for improving existing pricing frameworks.
References
[1] J. Andreasen. Stochastic volatility interest rate models. Encyclopedia of Quantitative Finance, 2010.
[2] D. Brigo. The general mixture-diffusion SDE and its relationship with an uncertain-volatility option model
with volatility-asset decorrelation. arXiv preprint arXiv:0812.4052, 2008.
[3] D. Brigo and F. Mercurio. Lognormal-mixture dynamics and calibration to market volatility smiles.
International Journal of Theoretical and Applied Finance, 5(04):427–446, 2002.
[4] D. Brigo and F. Mercurio. Interest rate models- theory and practice: with smile, inflation and credit.
Springer Finance, 2nd edition, 2007.
[5] R. Bulirsch and J. Stoer. Introduction to numerical analysis, volume 3. Springer, 2002.
[6] J. Casassus, P. Collin-Dufresne, and B. Goldstein. Unspanned stochastic volatility and fixed income
derivatives pricing. Journal of Banking & Finance, 29(11):2723–2749, 2005.
[7] D. Duffie, J. Pan, and K. Singleton. Transform analysis and asset pricing for affine jump-diffusions.
Econometrica, 68:1343–1376, 2000.
[8] H. G. Fong and O. A. Vašiček. Fixed-income volatility management. Journal of Portfolio Management,
17(4):41, 1991.
[9] D. Ga̧tarek and J. Jablecki. A nonparametric local volatility model for swaptions smile. Journal of
Computational Finance, 21(5):35–62, 2018.
[10] G. H. Golub and J. H. Welsch. Calculation of Gauss quadrature rules. Mathematics of Computation,
23(106):221–230, 1969.
[11] L. A. Grzelak. On randomization of affine diffusion processes with application to pricing of options on
VIX and S&P 500. arXiv:2208.12518, 2022.
[12] L. A. Grzelak, J. Witteveen, M. Suarez-Taboada, and C. W. Oosterlee. The stochastic collocation Monte
Carlo sampler: highly efficient sampling from expensive distributions. Quantitative Finance, 19(2):339–
356, 2019.
[13] D. Heath, R. Jarrow, and A. Morton. Bond pricing and the term structure of interest rates: A new
methodology for contingent claims valuation. Econometrica, 1(60):77–105, 1992.
[14] J. Hull and A. White. Pricing interest-rate derivative securities. The Review of Financial Studies, 3:573–
592, 1990.
[15] F. Jamshidian. An exact bond option formula. The Journal of Finance, 44:205–209, 1989.
20
[16] C. W. Oosterlee and L. A. Grzelak. Mathematical modeling and computation in finance. World Scientific,
2019.
[17] L. C. G. Rogers and D. Williams. Diffusions, Markov processes and martingales, volume 2 of Cambridge
Mathematical Library. Cambridge University Press, 2 edition, 2000.
[18] O. Vašiček. An equilibrium characterization of the term structure. Journal of Financial Economics,
5:177–188, 1977.
Table A.2: Selected distributions for the stochastic parameters. For the normal random variable
with some mean, µ, and variance, σ 2 , it is sufficient to consider standard normal distribution,
N (0, 1), and properly scale the θn points, obtained from Appendix A.2.
with R the support of ϑ, δi,j the Kronecker delta. Orthogonal polynomials {pi }N
i=0 satisfy the
following recursion relation:
where αi and βi are determined in terms of the moments of a random variable ϑ. Consider
the monomials mi (ϑ) = ϑi , and define µi,j as µi,j = E [mi (ϑ)mj (ϑ)], From all moments µi,j ,
we construct the Gram matrix M = {µi,j }N i,j=0 , which is symmetric and contains all moments
1 2N
{1, E[ϑ ], . . . , E[ϑ ]}. Since matrix M is, by definition, positive definite [10], we decompose
M = RT R, by the Cholesky decomposition of M.
The following step relates the Cholesky upper-triangular matrix R to the orthogonal poly-
nomials. This relationship has been established in [10] and is given by,
2
rj,j+1 rj−1,j rj+1,j+1
αj = − , j = 1, . . . , N, and βj = , j = 1, . . . , N − 1, (A.2)
rj,j rj−1,j−1 rj,j
with r0,0 = 1 and r0,1 = 0 and where ri,j is the (i, j)-th element of matrix R. Now, we
determine the symmetric tridiagonal matrix, J,
√
√α1 β1 √0 0 0
β1 α2 β2 0 0
. . .
J := .. .. .. ∈ RN ×N , (A.3)
p p
0 0 βN −2 p αN −1 βN −1
0 0 0 βN −1 αN
Randomization.
22